Liger-Kernel

Efficient Triton Kernels for LLM Training

Stars: 4142

Visit

Liger Kernel is a collection of Triton kernels designed for LLM training, increasing training throughput by 20% and reducing memory usage by 60%. It includes Hugging Face Compatible modules like RMSNorm, RoPE, SwiGLU, CrossEntropy, and FusedLinearCrossEntropy. The tool works with Flash Attention, PyTorch FSDP, and Microsoft DeepSpeed, aiming to enhance model efficiency and performance for researchers, ML practitioners, and curious novices.

README:

Liger Kernel: Efficient Triton Kernels for LLM Training

Stable		Nightly		Discord	Build

Supercharge Your Model with Liger Kernel

With one line of code, Liger Kernel can increase throughput by more than 20% and reduce memory usage by 60%, thereby enabling longer context lengths, larger batch sizes, and massive vocabularies.

Speed Up	Memory Reduction

Note:

Benchmark conditions: LLaMA 3-8B, Batch Size = 8, Data Type = bf16, Optimizer = AdamW, Gradient Checkpointing = True, Distributed Strategy = FSDP1 on 8 A100s.

Hugging Face models start to OOM at a 4K context length, whereas Hugging Face + Liger Kernel scales up to 16K.

Optimize Post Training with Liger Kernel

We provide optimized post training kernels like DPO, ORPO, SimPO, and more which can reduce memory usage by up to 80%. You can easily use them as python modules.

from liger_kernel.chunked_loss import LigerFusedLinearDPOLoss
orpo_loss = LigerFusedLinearORPOLoss()
y = orpo_loss(lm_head.weight, x, target)

Examples

Use Case	Description
Hugging Face Trainer	Train LLaMA 3-8B ~20% faster with over 40% memory reduction on Alpaca dataset using 4 A100s with FSDP
Lightning Trainer	Increase 15% throughput and reduce memory usage by 40% with LLaMA3-8B on MMLU dataset using 8 A100s with DeepSpeed ZeRO3
Medusa Multi-head LLM (Retraining Phase)	Reduce memory usage by 80% with 5 LM heads and improve throughput by 40% using 8 A100s with FSDP
Vision-Language Model SFT	Finetune Qwen2-VL on image-text data using 4 A100s with FSDP
Liger ORPO Trainer	Align Llama 3.2 using Liger ORPO Trainer with FSDP with 50% memory reduction

Key Features

Ease of use: Simply patch your Hugging Face model with one line of code, or compose your own model using our Liger Kernel modules.
Time and memory efficient: In the same spirit as Flash-Attn, but for layers like RMSNorm, RoPE, SwiGLU, and CrossEntropy! Increases multi-GPU training throughput by 20% and reduces memory usage by 60% with kernel fusion, in-place replacement, and chunking techniques.
Exact: Computation is exact—no approximations! Both forward and backward passes are implemented with rigorous unit tests and undergo convergence testing against training runs without Liger Kernel to ensure accuracy.
Lightweight: Liger Kernel has minimal dependencies, requiring only Torch and Triton—no extra libraries needed! Say goodbye to dependency headaches!
Multi-GPU supported: Compatible with multi-GPU setups (PyTorch FSDP, DeepSpeed, DDP, etc.).
Trainer Framework Integration: Axolotl, LLaMa-Factory, SFTTrainer, Hugging Face Trainer, SWIFT

Installation

Dependencies

CUDA

torch >= 2.1.2
triton >= 2.3.0

ROCm

torch >= 2.5.0 Install according to the instruction in Pytorch official webpage.
triton >= 3.0.0 Install from pypi. (e.g. pip install triton==3.0.0)

Optional Dependencies

transformers >= 4.x: Required if you plan to use the transformers models patching APIs. The specific model you are working will dictate the minimum version of transformers.

Note: Our kernels inherit the full spectrum of hardware compatibility offered by Triton.

To install the stable version:

$ pip install liger-kernel

To install the nightly version:

$ pip install liger-kernel-nightly

To install from source:

git clone https://github.com/linkedin/Liger-Kernel.git
cd Liger-Kernel

# Install Default Dependencies
# Setup.py will detect whether you are using AMD or NVIDIA
pip install -e .

# Setup Development Dependencies
pip install -e ".[dev]"

Getting Started

There are a couple of ways to apply Liger kernels, depending on the level of customization required.

1. Use AutoLigerKernelForCausalLM

Using the AutoLigerKernelForCausalLM is the simplest approach, as you don't have to import a model-specific patching API. If the model type is supported, the modeling code will be automatically patched using the default settings.

from liger_kernel.transformers import AutoLigerKernelForCausalLM

# This AutoModel wrapper class automatically monkey-patches the
# model with the optimized Liger kernels if the model is supported.
model = AutoLigerKernelForCausalLM.from_pretrained("path/to/some/model")

2. Apply Model-Specific Patching APIs

Using the patching APIs, you can swap Hugging Face models with optimized Liger Kernels.

import transformers
from liger_kernel.transformers import apply_liger_kernel_to_llama

# 1a. Adding this line automatically monkey-patches the model with the optimized Liger kernels
apply_liger_kernel_to_llama()

# 1b. You could alternatively specify exactly which kernels are applied
apply_liger_kernel_to_llama(
  rope=True,
  swiglu=True,
  cross_entropy=True,
  fused_linear_cross_entropy=False,
  rms_norm=False
)

# 2. Instantiate patched model
model = transformers.AutoModelForCausalLM("path/to/llama/model")

3. Compose Your Own Model

You can take individual kernels to compose your models.

from liger_kernel.transformers import LigerFusedLinearCrossEntropyLoss
import torch.nn as nn
import torch

model = nn.Linear(128, 256).cuda()

# fuses linear + cross entropy layers together and performs chunk-by-chunk computation to reduce memory
loss_fn = LigerFusedLinearCrossEntropyLoss()

input = torch.randn(4, 128, requires_grad=True, device="cuda")
target = torch.randint(256, (4, ), device="cuda")

loss = loss_fn(model.weight, input, target)
loss.backward()

High-level APIs

AutoModel

AutoModel Variant	API
AutoModelForCausalLM	`liger_kernel.transformers.AutoLigerKernelForCausalLM`

Patching

Model	API	Supported Operations
LLaMA 2 & 3	`liger_kernel.transformers.apply_liger_kernel_to_llama`	RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy
LLaMA 3.2-Vision	`liger_kernel.transformers.apply_liger_kernel_to_mllama`	RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy
Mistral	`liger_kernel.transformers.apply_liger_kernel_to_mistral`	RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy
Mixtral	`liger_kernel.transformers.apply_liger_kernel_to_mixtral`	RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy
Gemma1	`liger_kernel.transformers.apply_liger_kernel_to_gemma`	RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy
Gemma2	`liger_kernel.transformers.apply_liger_kernel_to_gemma2`	RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy
Qwen2, Qwen2.5, & QwQ	`liger_kernel.transformers.apply_liger_kernel_to_qwen2`	RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy
Qwen2-VL	`liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl`	RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy
Phi3 & Phi3.5	`liger_kernel.transformers.apply_liger_kernel_to_phi3`	RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy

Low-level APIs

Fused Linear kernels combine linear layers with losses, reducing memory usage by up to 80% - ideal for HBM-constrained workloads.
Other kernels use fusion and in-place techniques for memory and performance optimization.

Model Kernels

Kernel	API
RMSNorm	`liger_kernel.transformers.LigerRMSNorm`
LayerNorm	`liger_kernel.transformers.LigerLayerNorm`
RoPE	`liger_kernel.transformers.liger_rotary_pos_emb`
SwiGLU	`liger_kernel.transformers.LigerSwiGLUMLP`
GeGLU	`liger_kernel.transformers.LigerGEGLUMLP`
CrossEntropy	`liger_kernel.transformers.LigerCrossEntropyLoss`
Fused Linear CrossEntropy	`liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`

Alignment Kernels

Kernel	API
Fused Linear CPO Loss	`liger_kernel.chunked_loss.LigerFusedLinearCPOLoss`
Fused Linear DPO Loss	`liger_kernel.chunked_loss.LigerFusedLinearDPOLoss`
Fused Linear ORPO Loss	`liger_kernel.chunked_loss.LigerFusedLinearORPOLoss`
Fused Linear SimPO Loss	`liger_kernel.chunked_loss.LigerFusedLinearSimPOLoss`

Distillation Kernels

Kernel	API
KLDivergence	`liger_kernel.transformers.LigerKLDIVLoss`
JSD	`liger_kernel.transformers.LigerJSD`
Fused Linear JSD	`liger_kernel.transformers.LigerFusedLinearJSD`

Experimental Kernels

Kernel	API
Embedding	`liger_kernel.transformers.experimental.LigerEmbedding`
Matmul int2xint8	`liger_kernel.transformers.experimental.matmul`

Contributing, Acknowledgements, and License

Sponsorship and Collaboration

AMD: Providing AMD GPUs for our AMD CI.
Intel: Providing Intel GPUs for our Intel CI.
Modal: Free 3000 credits from GPU MODE IRL for our NVIDIA CI.
EmbeddedLLM: Making Liger Kernel run fast and stable on AMD.
HuggingFace: Integrating Liger Kernel into Hugging Face Transformers and TRL.
Lightning AI: Integrating Liger Kernel into Lightning Thunder.
Axolotl: Integrating Liger Kernel into Axolotl.
Llama-Factory: Integrating Liger Kernel into Llama-Factory.

Contact

For issues, create a Github ticket in this repository
For open discussion, join our discord channel
For formal collaboration, send an email to [email protected]

Cite this work

Biblatex entry:

@article{hsu2024ligerkernelefficienttriton,
      title={Liger Kernel: Efficient Triton Kernels for LLM Training},
      author={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},
      year={2024},
      eprint={2410.10989},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2410.10989},
      journal={arXiv preprint arXiv:2410.10989},
}

Star History

↑ Back to Top ↑

For Tasks:

Click tags to check more tools for each tasks

optimize training throughput reduce memory usage patch existing models compose custom models improve multi-gpu support

For Jobs:

machine learning engineer research scientist data scientist ai engineer deep learning specialist

Alternative AI tools for Liger-Kernel

Similar Open Source Tools

Liger-Kernel

github

: 4.1k

MooER

MooER (摩耳) is an LLM-based speech recognition and translation model developed by Moore Threads. It allows users to transcribe speech into text (ASR) and translate speech into other languages (AST) in an end-to-end manner. The model was trained using 5K hours of data and is now also available with an 80K hours version. MooER is the first LLM-based speech model trained and inferred using domestic GPUs. The repository includes pretrained models, inference code, and a Gradio demo for a better user experience.

github

: 124

agentscope

AgentScope is a multi-agent platform designed to empower developers to build multi-agent applications with large-scale models. It features three high-level capabilities: Easy-to-Use, High Robustness, and Actor-Based Distribution. AgentScope provides a list of `ModelWrapper` to support both local model services and third-party model APIs, including OpenAI API, DashScope API, Gemini API, and ollama. It also enables developers to rapidly deploy local model services using libraries such as ollama (CPU inference), Flask + Transformers, Flask + ModelScope, FastChat, and vllm. AgentScope supports various services, including Web Search, Data Query, Retrieval, Code Execution, File Operation, and Text Processing. Example applications include Conversation, Game, and Distribution. AgentScope is released under Apache License 2.0 and welcomes contributions.

github

: 5.8k

IDvs.MoRec

This repository contains the source code for the SIGIR 2023 paper 'Where to Go Next for Recommender Systems? ID- vs. Modality-based Recommender Models Revisited'. It provides resources for evaluating foundation, transferable, multi-modal, and LLM recommendation models, along with datasets, pre-trained models, and training strategies for IDRec and MoRec using in-batch debiased cross-entropy loss. The repository also offers large-scale datasets, code for SASRec with in-batch debias cross-entropy loss, and information on joining the lab for research opportunities.

github

: 119

skpro

skpro is a library for supervised probabilistic prediction in python. It provides `scikit-learn`-like, `scikit-base` compatible interfaces to: * tabular **supervised regressors for probabilistic prediction** \- interval, quantile and distribution predictions * tabular **probabilistic time-to-event and survival prediction** \- instance-individual survival distributions * **metrics to evaluate probabilistic predictions** , e.g., pinball loss, empirical coverage, CRPS, survival losses * **reductions** to turn `scikit-learn` regressors into probabilistic `skpro` regressors, such as bootstrap or conformal * building **pipelines and composite models** , including tuning via probabilistic performance metrics * symbolic **probability distributions** with value domain of `pandas.DataFrame`-s and `pandas`-like interface

github

: 231

dora

Dataflow-oriented robotic application (dora-rs) is a framework that makes creation of robotic applications fast and simple. Building a robotic application can be summed up as bringing together hardwares, algorithms, and AI models, and make them communicate with each others. At dora-rs, we try to: make integration of hardware and software easy by supporting Python, C, C++, and also ROS2. make communication low latency by using zero-copy Arrow messages. dora-rs is still experimental and you might experience bugs, but we're working very hard to make it stable as possible.

github

: 1.7k

llm4ad

LLM4AD is an open-source Python-based platform leveraging Large Language Models (LLMs) for Automatic Algorithm Design (AD). It provides unified interfaces for methods, tasks, and LLMs, along with features like evaluation acceleration, secure evaluation, logs, GUI support, and more. The platform was originally developed for optimization tasks but is versatile enough to be used in other areas such as machine learning, science discovery, game theory, and engineering design. It offers various search methods and algorithm design tasks across different domains. LLM4AD supports remote LLM API, local HuggingFace LLM deployment, and custom LLM interfaces. The project is licensed under the MIT License and welcomes contributions, collaborations, and issue reports.

github

: 123

cortex.cpp

Cortex is a C++ AI engine with a Docker-like command-line interface and client libraries. It supports running AI models using ONNX, TensorRT-LLM, and llama.cpp engines. Cortex can function as a standalone server or be integrated as a library. The tool provides support for various engines and models, allowing users to easily deploy and interact with AI models. It offers a range of CLI commands for managing models, embeddings, and engines, as well as a REST API for interacting with models. Cortex is designed to simplify the deployment and usage of AI models in C++ applications.

github

: 2.3k

openlit

OpenLIT is an OpenTelemetry-native GenAI and LLM Application Observability tool. It's designed to make the integration process of observability into GenAI projects as easy as pie – literally, with just **a single line of code**. Whether you're working with popular LLM Libraries such as OpenAI and HuggingFace or leveraging vector databases like ChromaDB, OpenLIT ensures your applications are monitored seamlessly, providing critical insights to improve performance and reliability.

github

: 1.0k

readme-ai

README-AI is a developer tool that auto-generates README.md files using a combination of data extraction and generative AI. It streamlines documentation creation and maintenance, enhancing developer productivity. This project aims to enable all skill levels, across all domains, to better understand, use, and contribute to open-source software. It offers flexible README generation, supports multiple large language models (LLMs), provides customizable output options, works with various programming languages and project types, and includes an offline mode for generating boilerplate README files without external API calls.

github

: 1.5k

qserve

QServe is a serving system designed for efficient and accurate Large Language Models (LLM) on GPUs with W4A8KV4 quantization. It achieves higher throughput compared to leading industry solutions, allowing users to achieve A100-level throughput on cheaper L40S GPUs. The system introduces the QoQ quantization algorithm with 4-bit weight, 8-bit activation, and 4-bit KV cache, addressing runtime overhead challenges. QServe improves serving throughput for various LLM models by implementing compute-aware weight reordering, register-level parallelism, and fused attention memory-bound techniques.

github

: 383

MaskLLM

MaskLLM is a learnable pruning method that establishes Semi-structured Sparsity in Large Language Models (LLMs) to reduce computational overhead during inference. It is scalable and benefits from larger training datasets. The tool provides examples for running MaskLLM with Megatron-LM, preparing LLaMA checkpoints, pre-tokenizing C4 data for Megatron, generating prior masks, training MaskLLM, and evaluating the model. It also includes instructions for exporting sparse models to Huggingface.

github

: 142

Starmoon

Starmoon is an affordable, compact AI-enabled device that can understand and respond to your emotions with empathy. It offers supportive conversations and personalized learning assistance. The device is cost-effective, voice-enabled, open-source, compact, and aims to reduce screen time. Users can assemble the device themselves using off-the-shelf components and deploy it locally for data privacy. Starmoon integrates various APIs for AI language models, speech-to-text, text-to-speech, and emotion intelligence. The hardware setup involves components like ESP32S3, microphone, amplifier, speaker, LED light, and button, along with software setup instructions for developers. The project also includes a web app, backend API, and background task dashboard for monitoring and management.

github

: 457

pixeltable

Pixeltable is a Python library designed for ML Engineers and Data Scientists to focus on exploration, modeling, and app development without the need to handle data plumbing. It provides a declarative interface for working with text, images, embeddings, and video, enabling users to store, transform, index, and iterate on data within a single table interface. Pixeltable is persistent, acting as a database unlike in-memory Python libraries such as Pandas. It offers features like data storage and versioning, combined data and model lineage, indexing, orchestration of multimodal workloads, incremental updates, and automatic production-ready code generation. The tool emphasizes transparency, reproducibility, cost-saving through incremental data changes, and seamless integration with existing Python code and libraries.

github

: 122

Q-Bench

Q-Bench is a benchmark for general-purpose foundation models on low-level vision, focusing on multi-modality LLMs performance. It includes three realms for low-level vision: perception, description, and assessment. The benchmark datasets LLVisionQA and LLDescribe are collected for perception and description tasks, with open submission-based evaluation. An abstract evaluation code is provided for assessment using public datasets. The tool can be used with the datasets API for single images and image pairs, allowing for automatic download and usage. Various tasks and evaluations are available for testing MLLMs on low-level vision tasks.

github

: 224

eko

Eko is a lightweight and flexible command-line tool for managing environment variables in your projects. It allows you to easily set, get, and delete environment variables for different environments, making it simple to manage configurations across development, staging, and production environments. With Eko, you can streamline your workflow and ensure consistency in your application settings without the need for complex setup or configuration files.

github

: 395

For similar tasks

Liger-Kernel

github

: 4.1k

verl

veRL is a flexible and efficient reinforcement learning training framework designed for large language models (LLMs). It allows easy extension of diverse RL algorithms, seamless integration with existing LLM infrastructures, and flexible device mapping. The framework achieves state-of-the-art throughput and efficient actor model resharding with 3D-HybridEngine. It supports popular HuggingFace models and is suitable for users working with PyTorch FSDP, Megatron-LM, and vLLM backends.

github

: 690

litdata

LitData is a tool designed for blazingly fast, distributed streaming of training data from any cloud storage. It allows users to transform and optimize data in cloud storage environments efficiently and intuitively, supporting various data types like images, text, video, audio, geo-spatial, and multimodal data. LitData integrates smoothly with frameworks such as LitGPT and PyTorch, enabling seamless streaming of data to multiple machines. Key features include multi-GPU/multi-node support, easy data mixing, pause & resume functionality, support for profiling, memory footprint reduction, cache size configuration, and on-prem optimizations. The tool also provides benchmarks for measuring streaming speed and conversion efficiency, along with runnable templates for different data types. LitData enables infinite cloud data processing by utilizing the Lightning.ai platform to scale data processing with optimized machines.

github

: 395

llm-awq

AWQ (Activation-aware Weight Quantization) is a tool designed for efficient and accurate low-bit weight quantization (INT3/4) for Large Language Models (LLMs). It supports instruction-tuned models and multi-modal LMs, providing features such as AWQ search for accurate quantization, pre-computed AWQ model zoo for various LLMs, memory-efficient 4-bit linear in PyTorch, and efficient CUDA kernel implementation for fast inference. The tool enables users to run large models on resource-constrained edge platforms, delivering more efficient responses with LLM/VLM chatbots through 4-bit inference.

github

: 2.6k

duo-attention

DuoAttention is a framework designed to optimize long-context large language models (LLMs) by reducing memory and latency during inference without compromising their long-context abilities. It introduces a concept of Retrieval Heads and Streaming Heads to efficiently manage attention across tokens. By applying a full Key and Value (KV) cache to retrieval heads and a lightweight, constant-length KV cache to streaming heads, DuoAttention achieves significant reductions in memory usage and decoding time for LLMs. The framework uses an optimization-based algorithm with synthetic data to accurately identify retrieval heads, enabling efficient inference with minimal accuracy loss compared to full attention. DuoAttention also supports quantization techniques for further memory optimization, allowing for decoding of up to 3.3 million tokens on a single GPU.

github

: 57

For similar jobs

sweep

Sweep is an AI junior developer that turns bugs and feature requests into code changes. It automatically handles developer experience improvements like adding type hints and improving test coverage.

github

: 7.1k

teams-ai

The Teams AI Library is a software development kit (SDK) that helps developers create bots that can interact with Teams and Microsoft 365 applications. It is built on top of the Bot Framework SDK and simplifies the process of developing bots that interact with Teams' artificial intelligence capabilities. The SDK is available for JavaScript/TypeScript, .NET, and Python.

github

: 390

ai-guide

This guide is dedicated to Large Language Models (LLMs) that you can run on your home computer. It assumes your PC is a lower-end, non-gaming setup.

github

: 159

classifai

Supercharge WordPress Content Workflows and Engagement with Artificial Intelligence. Tap into leading cloud-based services like OpenAI, Microsoft Azure AI, Google Gemini and IBM Watson to augment your WordPress-powered websites. Publish content faster while improving SEO performance and increasing audience engagement. ClassifAI integrates Artificial Intelligence and Machine Learning technologies to lighten your workload and eliminate tedious tasks, giving you more time to create original content that matters.

github

: 593

chatbot-ui

Chatbot UI is an open-source AI chat app that allows users to create and deploy their own AI chatbots. It is easy to use and can be customized to fit any need. Chatbot UI is perfect for businesses, developers, and anyone who wants to create a chatbot.

github

: 27.7k

BricksLLM

BricksLLM is a cloud native AI gateway written in Go. Currently, it provides native support for OpenAI, Anthropic, Azure OpenAI and vLLM. BricksLLM aims to provide enterprise level infrastructure that can power any LLM production use cases. Here are some use cases for BricksLLM: * Set LLM usage limits for users on different pricing tiers * Track LLM usage on a per user and per organization basis * Block or redact requests containing PIIs * Improve LLM reliability with failovers, retries and caching * Distribute API keys with rate limits and cost limits for internal development/production use cases * Distribute API keys with rate limits and cost limits for students

github

: 953

uAgents

uAgents is a Python library developed by Fetch.ai that allows for the creation of autonomous AI agents. These agents can perform various tasks on a schedule or take action on various events. uAgents are easy to create and manage, and they are connected to a fast-growing network of other uAgents. They are also secure, with cryptographically secured messages and wallets.

github

: 1.2k

griptape

Griptape is a modular Python framework for building AI-powered applications that securely connect to your enterprise data and APIs. It offers developers the ability to maintain control and flexibility at every step. Griptape's core components include Structures (Agents, Pipelines, and Workflows), Tasks, Tools, Memory (Conversation Memory, Task Memory, and Meta Memory), Drivers (Prompt and Embedding Drivers, Vector Store Drivers, Image Generation Drivers, Image Query Drivers, SQL Drivers, Web Scraper Drivers, and Conversation Memory Drivers), Engines (Query Engines, Extraction Engines, Summary Engines, Image Generation Engines, and Image Query Engines), and additional components (Rulesets, Loaders, Artifacts, Chunkers, and Tokenizers). Griptape enables developers to create AI-powered applications with ease and efficiency.

github

: 2.1k